| Ultrasound detection is a widely used method for cardiac medical examination,offering the benefits of being inexpensive,radiation-free,and non-invasive.In the quantitative cardiac analysis based on echocardiography,the localization of cardiac anatomical landmarks is a critical step.However,manual labeling of these landmarks by professional physicians is a costly,time-consuming,and labor-intensive process that is also prone to both inter-observer and intra-observer variations.Therefore,it is necessary to design an algorithm to realize the automatic localization of cardiac anatomical landmarks.This thesis focuses on developing an algorithm for localizing anatomical landmarks in the right ventricle.The main research contributions can be summarized in the following three points:(1)A heatmap-regression algorithm for localizing anatomical landmarks of the right ventricle is proposed.In this thesis,the U-Net model serves as the foundation for the algorithm,with two modifications made to enhance its performance.Firstly,we use Res Net to improve the encoding part of the U-Net model,which enhances the feature extraction ability of the model.Secondly,a Coord Conv layer is added after the input layer of the U-Net network,which enhances the model’s ability to extract the coordinate relationship of landmarks.Moreover,this thesis analyzes the advantages and limitations of three different loss functions,namely L2 Loss,BCE Loss,and Adaptive Wing Loss,in right ventricle anatomical landmark localization.Building on this analysis,the thesis proposes two enhancements to the loss function.Firstly,we regress the value of the previous layer before the Sigmoid activation layer during the training process,helping train the model more effectively.Secondly,Weighted Smooth L1 Loss and Focal Smooth L1 Loss are introduced as two new loss functions that not only balance foreground and background points but are also more robust to noise.(2)A multi-task framework simultaneously performing right ventricle anatomical landmark localization and echocardiographic view classification is proposed.Typically,the echocardiographic view classification task is conducted as an upstream task of the right ventricle anatomical landmark localization task.In this multi-task framework,the two tasks’ overall parameter amount and computation load are reduced by sharing the Res Net network in the encoding layer.Additionally,a new attention module is introduced to allow the view classification task to provide channel-wise attention and spatial attention to the landmark localization task.Compared with the framework for two separate tasks,the multi-task framework has lower parameters and computation requirements as well as faster inference speed.Remarkably,the performance of the proposed multi-task framework is comparable to that of the framework of two separate tasks.This thesis divides the training process of the multi-task framework into two stages based on the involvement of the attention module.In the first stage,the attention module is not included in the training process.During both training stages,either one of the task branches can be trained in each epoch by randomly selecting branches according to the probability,or two task branches can be trained simultaneously through the multi-task loss function.Furthermore,this thesis examines the impact of multi-frame integration on the landmark localization effect of the multi-task model.(3)An echocardiography-oriented anatomical landmark labeling system for the right ventricle is implemented.This thesis designs a set of basic procedures for labeling the anatomical landmarks of the right ventricle on the echocardiography for this system,which provides a reference for the quantitative analysis of the heart based on echocardiography.Moreover,by applying the multi-task model proposed in this thesis,the labeling system can help physicians to label the right ventricle anatomical landmarks on the echocardiography more conveniently and efficiently. |
PDF Full Text Request